1. Operational forecasts of strong precipitation events over complex topography Pierre Eckert Ralf Kretschmar Christof Appenzeller MétéoSuisse

2. Motivation: model bias, downscaling Three techniques: EFI, LEPS, ANN EFI ANN (Artificial Neural Networks) November 2002 case Conclusions Operational forecasts of strong precipitation events over complex topography

3. Probability forecasts Usually based on an Ensemble Prediction System (EPS) DMO Statistical interpretation … Methodes usually not well suited for rare events: Positive skill scores for prob(rr>10 mm/24h) Negative skill scores for prob(rr>50 mm/24h)

4. ECMWF Medium range model “does” generate severe weather events: forecast/observed frequency ratio (FBI) EFI LEPS

5. Three „rescaling / downscaling“ techniques for extreme events Extreme Forecast Index (EFI)  rescale with respect to model climate LEPS  downscale ensemble with a LAM Artificial neural Network (ANN)  pattern recognition of extreme situations with respect to a given meteorological parameter (precipitation)

6. p F f (p) EFI: Extreme Forecast Index (ECMWF)

7. Verification parameters  Precipitations (nov.2001-apr.2003) Predictands: 2 nd max prec. / 24h of climatological regions (W, E, S) Predictors: EFI max over regions  Thresholds 10 mm/24h 20 mm/24h 50 mm/24h

8. Verification: scores EventObservedNot Observed ForecastedAB Not ForecastedCD Forecasted : Index / prob > given threshold Hit Rate = A / (A+C) FAR = B / (A+B)

9. EFI verification: Switzerland South

10. Recognition of extreme events with ANN ANN: Artificial Neural Network Non linear classification method Unsupervised vs supervised training

11. Unsuprevised training Self Organising Map (SOM) = classification of weather patterns irrespective of the predictand (precipitation) (unsupervised training) Interpretation in terms of weather elements made in a second step: probabilty to exceed some threshold for each element of the classification

12. Suprevised training The predictors and the predictand (precipitation) have to be shown simultaneously to the network (supervised training) Feed Foreward Neural Network (FFNN) shown here

13. FFNN: Feed Foreward Neural Network

14. Predictors

15. Predictand 24 hour rainfall in Lugano (south of the Alps) Q95: 95% quantile, 20 day return period Q95=28mm Q99: 99% quantile, 100 day return period Q99=58mm

16. SOM (unsupervised) FFNN (supervised) ANN upper air Rainfall Lugano

17. SOM (unsupervised) FFNN (supervised) ANN upper air Rainfall Lugano

18. What about DMO precipitation? Use ECMWF T511 DMO precipitation Simple rescaling: Event Q95 predicted when DMO rr>10mm, 20mm, 30mm,… Use FFNN output from upper air fields + DMO rainfall to train a new ANN

19. Rescaled DMO (T511) FFNN + precip DMO ANN upper air + precip DMO Rainfall Lugano

20. Rescaled DMO (T511) FFNN + precip DMO ANN upper air + precip DMO Rainfall Lugano

21. November 2002 flooding 14-16th November Low over western Mediterranean Southerly current over the Alps Over 100 mm/day south of the Alps (classical) Geneva: 92 mm in one day (14th)  Very exceptional

22. TP050 RRSS Niederschlag; Tagessumme (konv.Periode: 0540-0540 Folgetag) (0.1 mm)................................................................................ SHA 119 BAS 304 GUT 48 RUE 246 LAE // TAE 70 FAH 474 KLO 98 STG 66 BUS 152 REH 94 SMA 40 HOE // CHA 245 WYN 230 WAE 45 SAE 105 CDF 544 NAP 106 LUZ 52 VAD 171 NEU 355 BER 226 PIL 64 GLA 195 FRE 651 PAY 309 ALT 246 CHU 347 PLF 191 INT 45 ENG 113 WFJ 128 SCU 504 MLS 51 JUN // GUE 72 DIS 794 DAV 225 DOL 766 PUY 468 ABO 100 GRH 243 PIO 1123 HIR 1765 SAM 767 CGI 870 AIG 132 MVE 59 ULR 572 ROE 1101COM 1041SBE 1359 COV 530 GVE 926 SIO 65 VIS 40 CIM 792 ROB 800 FEY 92 OTL 1230MAG 2037 EVO 24 ZER 164 LUG 637 GSB 906 SBO 710................................................................................ 14.11.2002

23. DMO

26. LEPS rr>20 mm, 14.11.2002 12z – 15.11.2002 12z LUG ZUR SIO GVE VAD

27. LEPS rr>50 mm, 14.11.2002 12z – 15.11.2002 12z LUG ZUR SIO GVE VAD

28. LEPS rr>100 mm, 14.11.2002 12z – 15.11.2002 12z LUG ZUR SIO GVE VAD

29. LEPS rr>150 mm, 14.11.2002 12z – 15.11.2002 12z LUG ZUR SIO GVE VAD

30. Conclusions Forecasting of rare events requests special treatment Downscaling / rescaling Hit rate is usable False Alarm Rate can be optimised DMO, EFI, LEPS, ANN have each their own qualities Define methods based on combinations ot these techiques Critical eye of the forecaster is still requested, mostly for reducing the False Alarm Rate.